Mapping activity locations of people from social media data enriched with time-use and satellite data: the case of cross-border commuting to Luxembourg

Mobile big data, including social media data and other data collected from mobile devices, are widely used to study the mobility and spatial behaviour of people in dynamic societies. A key challenge in social media-based research is detecting activity locations, often hindered by sparse data and limited user representativeness. Existing methods typically rely on simplistic assumptions, neglecting ancillary data sources to improve modelling. We propose a framework that integrates time and land use information to enhance activity location detection. Using geolocated Twitter archives (2010–2020), enriched with satellite and time-use survey data, we detected home and work locations to analyse cross-border commuters in Luxembourg's Greater Region. Evaluating 100 model variations, we compared individual-level expert assessments with aggregated municipality-level official statistics, revealing notable impacts of the models' parameters on prediction accuracy. The best-performing model employed a fuzzy combination of satellite-based and time-use survey probabilities, yielding an F1 score of around 0.8 at the country level and 0.4 at the municipality level for work and home prediction. The model outperformed the widely-used OSNA model by 2.5–7 times. We discuss the implications of our framework and suggest pathways for improving the applicability of various mobile big data sources in the future.